A Fast Weighted Fuzzy C-Medoids Clustering for Time Series Data Based on P-Splines.

Area of Science:

• Data Science
• Machine Learning
• Time Series Analysis

Background:

• Digital information growth generates massive, often noisy, time series data.
• Outliers in time series data degrade the effectiveness of clustering algorithms.
• Efficiently extracting statistical insights from complex time series is challenging.

Purpose of the Study:

• To propose a novel clustering algorithm for time series datasets.
• To enhance the robustness and efficiency of time series clustering.
• To address the challenges posed by noisy data and outliers.

Main Methods:

• Developed a fast weighted fuzzy C-medoids clustering algorithm based on P-splines (PS-WFCMdd).
• Utilized P-splines for fitting functional data from original time series, creating smooth inputs.
• Incorporated a new weighted method to mitigate outlier influence and improve robustness.
• Employed the Mueen's Algorithm for Similarity Search (version 3) for efficient similarity measurement.

Main Results:

• The PS-WFCMdd algorithm demonstrates improved processing speed for time series data.
• Experimental evaluations show good comprehensive performance across various clustering metrics.
• The method effectively handles noisy data and outliers, enhancing clustering quality.

Conclusions:

• The proposed PS-WFCMdd algorithm offers an effective solution for time series clustering.
• P-spline fitting and a novel weighting strategy significantly improve robustness and efficiency.
• The algorithm provides a valuable tool for discovering hidden statistical information in large time series datasets.